jeff-adapter-code-router-gguf

The LoRA GGUF of jeff-adapter-code-router (Jeff v1.3), for llama.cpp.

Jeff-Code router: should the coding model think hard on this turn?. In the Jeff-Code agent, picks how hard Qwen3.8-27B should think on each turn (off, low, medium, xhigh), so easy turns run fast.

This repository holds only the adapter. Load it on top of the base GGUF from mstrasser/jeff-base-gguf; there is no separate full model per adapter.

Serve code-router on the Q8_0 base only. The Jeff-Code router's choices are close calls, so at Q4_K_M it changes about 6% of its thinking on/off decisions (Q8_0: 0.8%). Use models/base-q8_0.gguf, not models/base-q4_k_m.gguf.

Files

code-router.jeff.json has a temperature for each format (temperature_by_format: f16, q8_0, q4_k_m); use the one for your base format, so the probabilities stay calibrated.

Results

The same rows, through llama.cpp: the base GGUF plus this LoRA GGUF, with the temperature refitted for each format. Running Jeff with llama.cpp

Run-time rule: the same thinking on/off decision (on when P(xhigh) ≥ 0.6). The GGUF takes the same decision as full precision on 99.2% of the 987 development rows at Q8_0 and 94.3% at Q4_K_M.

Full precision: full precision from the trainer's own evaluation on the development split.

More on the adapter's card: mstrasser/jeff-adapter-code-router.

How to run it

Use llama.cpp commit cb7934c52ca8710994b2ecc19775ebefcfdb8d01 or newer: it needs the qwen35 architecture and LoRA on the output layer.

hf download mstrasser/jeff-base-gguf --local-dir jeff-gguf
hf download mstrasser/jeff-adapter-code-router-gguf --local-dir jeff-gguf
cd jeff-gguf
llama-server -m models/base-q8_0.gguf -c 8192 -np 1 --lora-init-without-apply --lora loras/code-router.gguf

Jeff is not a chat model. For each decision you run one forward pass over the prompt and read the probabilities of the answer codes; you never sample text.

1. Build the prompt exactly as Jeff does (the question, the state, the options as answer codes, and the changing state field under "Latest", then the chat template with thinking off). The Jeff repository builds it for you (jeff.model.decision_messages).
2. Tokenize without a beginning-of-sequence token, with special tokens parsed.
3. Run one forward pass over the whole prompt from an empty state, with this adapter active. Clear the state between prompts: Qwen3.5 has recurrent layers, so a reused state would carry over.
4. Read the probabilities. Take the logits of the first N answer-code tokens (token_ids[:N], N = the number of options), divide by the temperature for this adapter and your format, and apply a softmax over those N.

With llama-server, each decision is one POST /completion with "n_predict": 1, "cache_prompt": false, a lora list that names every loaded adapter (scale 1 for this one, 0 for the rest; an adapter left out of the list keeps scale 1.0), a logit bias of +1000 on every answer-code token of the request, and n_probs set to the number of options. From the llama.cpp library: load the base once with llama_model_load_from_file, this adapter once with llama_adapter_lora_init, and per request call llama_set_adapters_lora with scale 1, clear the memory, decode once and read llama_get_logits_ith(ctx, -1).

The full guide, with an example prompt and request: Running Jeff with llama.cpp.

Data and licence

Adapter licence: Apache-2.0.

Qwen3.5-0.8B notice: these weights were modified from Qwen3.5-0.8B by the Jeff project: jeff-base is a fine-tune of Qwen3.5-0.8B, and this adapter was trained on top of it. Qwen3.5-0.8B is Copyright 2026 Alibaba Cloud and licensed under the Apache License, Version 2.0; a copy of that licence is in LICENSE.

To confirm: that the licences of the session data and of the benchmark tasks allow training and publishing the adapter.

It was trained on:

• Recorded Qwen3.8-27B sessions (at xhigh), re-asked at lower thinking levels. Licence: To confirm (licence not confirmed yet) · Made by Qwen3.8-27B (the re-asked answers); Qwen3.8-Max (hosted, Alibaba Cloud DashScope), for the yes-or-no judgements the code rule could not make

  Sessions recorded at medium thinking were left out. The re-asked answers come from Qwen3.8-27B; where the code rule could not decide, Qwen3.8-Max judged whether the lower-level step was as good.

  To confirm: the licence of the source sessions

Links

• The adapter at full precision: mstrasser/jeff-adapter-code-router (revision v1.3)
• The base GGUF: mstrasser/jeff-base-gguf
• Adapter card: mstrasser/jeff-adapter-code-router
• Running Jeff with llama.cpp: jeffhub.ai/docs/llama-cpp

Jeff is an independent project. It uses the same request format as Jev but is not affiliated with or endorsed by TypeSafe, the makers of Jev.